Regional power grid new energy consumption dynamic prediction system

Abstract

The invention discloses a dynamic prediction system for new energy consumption of a regional power grid, and relates to the field of power systems. The feature fusion unit is used for extracting spatial distribution features of input data; the BiLSTM unit is used for mining time correlation of an input data sequence; and the AM unit adaptively adjusts the weight of the BiLSTM model for each input data calculation value according to the data sequence. And powerful intelligent support can be provided for power grid dispatching optimization, reserve capacity configuration and market transaction under the background of high-proportion access of new energy.

Description

Technical Field

[0001] This invention relates to the field of power systems, and more particularly to a dynamic prediction system for the consumption of new energy sources in a regional power grid. Background Technology

[0002] Renewable energy sources such as wind power and solar power are rapidly replacing traditional fossil fuels due to their clean and low-carbon advantages, becoming an important pillar of my country's energy structure transformation. However, the output of these new energy sources is characterized by significant intermittency, strong volatility, and high uncertainty, disrupting the traditional power system's operation mode of source-load-driven and stable regulation, leading to a fundamental change in the dynamic characteristics of the power system. With the rapid increase in the penetration rate of new energy sources, the difficulty of predicting power load and power output has increased dramatically. This not only weakens the grid's frequency regulation and peak-shaving capabilities but also significantly increases the demand for spinning reserve and reserve capacity, triggering a series of systemic operational risks and severely restricting the safe and efficient consumption of new energy sources and the stable operation of the power grid.

[0003] Current forecasting technologies exhibit multiple shortcomings when addressing scenarios with a high proportion of renewable energy. First, traditional time-series forecasting methods (such as ARIMA and unidirectional LSTM) are generally based on linear assumptions or single-dimensional data input, making it difficult to characterize the complex nonlinear mapping relationships between wind and solar power output and multi-source heterogeneous variables such as meteorological factors, equipment operating status, and electricity price fluctuations. This results in limited forecasting accuracy and stability. Second, cross-regional renewable energy forecasting tasks rely on techniques like transfer learning for model generalization. However, neglecting the dynamic changes in data distribution due to regional geographical and climatic differences often leads to performance degradation after model transfer, and even negative transfer, severely impacting its practicality and reliability. Furthermore, in actual grid operation, renewable energy power fluctuates dramatically and changes rapidly, requiring forecasting systems to have dynamic update capabilities at the minute or even second level. Current deep learning forecasting models typically suffer from high computational redundancy and prolonged response times, making it difficult to meet the high real-time requirements of grid regulation. More critically, most current forecasting systems lack an efficient linkage mechanism with grid dispatching platforms. The inability to provide real-time feedback to the dispatching system to support decision-making breaks the closed-loop management chain of "prediction-control-feedback," making it difficult for theoretical models to effectively guide grid dispatching practices. This "information silo" phenomenon is particularly prominent in the context of high-proportion renewable energy grid integration, leading to lagging dispatching strategies and a disconnect between power trading mechanisms and prediction results, further impacting the economic viability of renewable energy consumption and the security of regulation. Summary of the Invention

[0004] In view of the above-mentioned deficiencies of the prior art, the present invention provides a dynamic prediction system for the consumption of new energy in a regional power grid, comprising:

[0005] The eigenvalue decomposition unit decomposes VMP data and photovoltaic power generation data into multiple scales through wavelet transform to obtain time series under each frequency component;

[0006] The feature fusion unit is used to extract the spatial distribution features of the input data;

[0007] BiLSTM units are used to mine the temporal correlation of input data sequences;

[0008] The AM unit adaptively adjusts the weights of the BiLSTM model for each input data value according to the data order.

[0009] Furthermore, the feature fusion unit employs a convolutional neural network.

[0010] Furthermore, the BiLSTM unit includes an input layer, forward and reverse LSTM layers, forward and reverse LSTM layers, and an output layer.

[0011] Furthermore, the weights of the attention unit are:

[0012]

[0013] Among them, e t For Q t The attention probability allocation value is determined at time t, where u and w are weighting coefficients and b is the bias coefficient.

[0014] Furthermore, Q t The attention probability allocation value e determined at time t t for:

[0015] e t =utanh(wQ) t +b).

[0016] Compared with the prior art, the present invention has the following technical effects:

[0017] This invention, with "multi-source fusion, dynamic learning, real-time response, and deep coupling" as its core design principles, innovatively constructs a multi-dimensional heterogeneous data fusion mechanism. It integrates key factors such as meteorological cloud maps, load history, power grid topology, electricity price signals, and equipment status. Through spatiotemporal feature joint modeling, it comprehensively reveals the high-order coupling relationship between renewable energy output and the system environment. The main body of the model employs a bidirectional long short-term memory network (BiLSTM) to fully extract the bidirectional dependence features of renewable energy in the time dimension, effectively enhancing the ability to perceive future trend changes. An attention mechanism is introduced to achieve dynamic weight adjustment of input features, improving the model's ability to focus on key variables and significantly optimizing prediction accuracy.

[0018] The proposed regional power grid renewable energy dynamic prediction system based on spatiotemporal feature fusion and BiLSTM-AM overcomes the technical bottlenecks of traditional methods in nonlinear modeling, cross-domain generalization, real-time response, and system coupling. It provides strong intelligent support for grid dispatch optimization, reserve capacity allocation, and market trading under the background of high renewable energy integration. This achievement provides a key technical path for building a new type of intelligent power system and promoting the clean and low-carbon transformation of energy, and has significant theoretical value and engineering application implications.

[0019] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of a specific embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the wavelet transform principle of the present invention with reference to a specific embodiment;

[0022] Figure 3 This is a CNN structure diagram of a specific embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the structure of a BiLSTM unit according to a specific embodiment of the present invention;

[0024] Figure 5 This is a model structure of the attention mechanism unit in a specific embodiment of the present invention. Detailed Implementation

[0025] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0026] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0027] Some exemplary embodiments of the invention have been described for illustrative purposes. It should be understood that the invention may be implemented in other ways not specifically shown in the accompanying drawings.

[0028] like Figure 1 As shown, a regional power grid renewable energy consumption dynamic prediction system includes:

[0029] The eigenvalue decomposition unit uses wavelet transform to decompose VMP data and photovoltaic power generation data into multi-scale data to obtain time series data for each frequency component.

[0030] Wavelet transform (WT) has significant advantages in time-frequency localization. Compared to the traditional Fourier transform, which only provides frequency domain information, WT can extract both time and frequency domain features simultaneously, thus more comprehensively reflecting the changing patterns of new energy power generation data. In this embodiment, wavelet analysis is used to decompose the data into wavelets of different scales and operating frequencies. These wavelets can better describe the details and local features of the received signal, revealing the implicit periodic oscillations in the time series. By using wavelet decomposition to decompose the power load signal into different frequency components, the details and local features of the signal can be observed, and some noise in the signal can be removed. The wavelet function used in this embodiment is:

[0031]

[0032] Where a0 is the scaling parameter, b0 is the translation parameter, m is the scaling constant (an integer), n is the translation constant (an integer), x(t) is the time series, and ψx(x) is the complex conjugate function.

[0033] Perform a discrete wavelet transform on the input data to obtain the coefficient matrix of the wavelet decomposition [CA]. n CD n CD n-1 [, ..., CD2, CD1], where element CA n The low-frequency components represent approximate components of the data, while the elements CD…CD are high-frequency components, representing detail components. This paper decomposes the load data into two layers: the load S is decomposed into detail components CD and approximate components CA. CA is further decomposed into detail components CD and another approximate component CA, as detailed below. Figure 2 As shown.

[0034] In this embodiment, wavelet transform is used to perform multi-scale decomposition on the original NWP data and photovoltaic power generation data to obtain time series of different frequency components. Specifically, the information in the frequency range of 0–120Hz of the original photovoltaic power data is decomposed into several sub-frequency band data sequences using wavelet transform. Here, A1 represents the low-frequency subsequence with a frequency range of 0–4Hz, mainly reflecting the long-term trend characteristics of the original data; A2 to A5 are subsequences with corresponding frequencies of 4–8Hz, 8–16Hz, 16–32Hz, and 32–64Hz, respectively, used to capture the mid-to-high frequency fluctuations and short-term randomness present in the original data. The wavelet transform method can effectively reduce the non-stationarity and complexity of the original data, enabling the model to learn in a targeted manner for different scale features. The decomposed multi-scale sequences are used as inputs to the prediction model for modeling and calculation, and finally, the prediction results of each frequency level are fused to obtain the global prediction value for the output of new energy power generation. This method not only preserves the important time-related characteristics of the original data but also achieves in-depth mining of its volatility and randomness, significantly improving the generalization ability and accuracy of the prediction model. By introducing wavelet transform into the photovoltaic power and meteorological data processing flow, and combining it with subsequent spatiotemporal feature fusion and attention mechanism-optimized bidirectional long short-term memory network (BiLSTM-AM), this invention demonstrates good application potential in improving the accuracy of new energy power generation prediction, and provides more reliable dynamic prediction support for the consumption of new energy in regional power grids.

[0035] The feature fusion unit is used to extract the spatial distribution features of the input data; in this embodiment, the feature fusion unit adopts a convolutional neural network, and its structure is as follows: Figure 3 As shown, Convolutional Neural Networks (CNNs), through local connectivity, weight sharing, and pooling, effectively reduce the number of parameters and data dimensionality of deep neural networks while deeply mining the spatial features of input data, thus significantly improving the computational speed and data analysis capabilities of CNNs. The basic structure of a CNN consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer.

[0036] The input layer is used to acquire the input data for the CNN. In this application, the input data consists of photovoltaic power generation data and NWP data. When the unit scale of the input data is inconsistent, it will seriously affect the prediction accuracy of the model. In order to eliminate the impact of inconsistent unit scale on the prediction accuracy of the model, it is necessary to normalize the data before transmitting it to the input layer. In this application, the photovoltaic power generation data and NWP data are normalized to the range of [0,1]. The input data is an N*M data matrix, where N is the time series length of the data, and M is the data information of power, wind speed, wind direction, temperature, and humidity at each time point.

[0037] The convolutional layer is the core component of CNN, which uses convolution operations to extract the spatial distribution features of the input data. In order to extract the seasonal and correlation features between photovoltaic power generation data and NWP data, this embodiment uses 50 convolution kernels to convolve the input data. The size of the convolution kernel is 3*3 and the stride is 1. The convolution calculation formulas are given by equations (1) and (2) respectively.

[0038]

[0039]

[0040] In equations (1) and (2) above, K d If it is the d-th convolutional kernel, then the kernel size matrix is ​​l×l, f is the activation function of the neuron, and b d It is the threshold of the d-th activation function. The outputs of the neurons in the i-th row and j-th column are calculated for the d-th convolutional kernel.

[0041] The linear correction function is simple to calculate and can effectively overcome the gradient vanishing problem of activation functions such as tanh and logistic functions. Therefore, a rectified linear unit (ReLU) is used as the activation function of the neuron in the convolution calculation, and the linear rectified function is shown in equation (3):

[0042] f(x) =ReLU(x) = max(0, 1) (3)

[0043] The calculation process of convolutional layers is as follows: Figure 3 As shown in the convolutional layer.

[0044] The pooling layer is used to reduce the dimensionality of the input high-dimensional feature data, improve the computation speed of CNN, and prevent overfitting. Commonly used pooling calculation methods include max pooling, mean pooling, and random pooling. In this embodiment, based on the distribution characteristics of photovoltaic power generation data and NWP data, max pooling is selected for calculation, with a pooling kernel size of 2*2 and a step size of 2; the max pooling calculation process of the data is shown in equation (4).

[0045] z ij = max(y i,j , y i,j+1 , y i+1,j , y i+1,j+1 (4)

[0046] In equation (4), y represents the feature data output by the convolutional layer, and z... ij This is the result of max pooling. The calculation principle of max pooling is as follows: Figure 2 The pooling layer is shown in the diagram.

[0047] The fully connected layer connects each neuron to all neurons in the previous layer and combines local features from the convolutional or pooling layers into global features. In this embodiment, the fully connected layer adopts a double fully connected layer structure. The first fully connected layer merges local features from the pooling layer into global features. The second fully connected layer serves as the regression output layer, outputting a series of feature data. The fully connected computation of the neural network is shown in Equation (5).

[0048] y z =f(w z x z-1 +b z (5)

[0049] In equation (5), y z x is the output of the z-th layer neuron. z-1 For the output of layer (-1), w z and b z Let be the weights and biases of the fully connected layer z. In this embodiment, the last two layers of the CNN are fully connected layers.

[0050] The output layer is used to output seasonal and correlation features between data, and extracts seasonal and correlation features between data from photovoltaic power generation data and NWP data through CNN.

[0051] BiLSTM units are used to mine the temporal correlation of input data sequences;

[0052] like Figure 4 As shown, the Bidirectional Long Short-Term Memory (BiLSTM) model in this embodiment consists of an input layer, a forward LSTM, a backward LSTM, and an output layer. The forward LSTM and backward LSTM can obtain past and future information of the input time series data, respectively, and the output Qt is obtained from the hidden states of the forward LSTM. The hidden state of the inverse LSTM at time t composition.

[0053] LSTM models can effectively extract forward data features from temporal data. However, temporal data is related not only to previous data but also to future data. To explore the forward and backward relationships between temporal data in depth, this embodiment employs a bidirectional long short-term memory (BiLSTM) unit, including an input layer, forward and backward LSTM layers, forward and backward LSTM layers, and an output layer, as shown in the structure below. Figure 4 As shown, the forward LSTM layer and the backward LSTM layer can obtain past and future information of the input time series data, respectively. The output Qt is determined by the hidden states of the forward LSTM. The hidden state of the inverse LSTM at time t composition.

[0054] Hidden state of normal LSTM Hidden state of inverse LSTM The output layer Qt can be calculated using equations (6)-(8).

[0055]

[0056] Where σ is the activation function, b is the bias vector, and W xsf W is the weight matrix of the input layer. sfsf This is the weight matrix of the hidden layer in a forward LSTM. This is the weight matrix of the hidden layer in the inverse LSTM. and These are the weight matrices of the output layer.

[0057] The AM unit is used to adaptively adjust the weights of the BiLSTM model for each input data value according to the data order.

[0058] Attention mechanism (AM) is a resource allocation mechanism that mimics the human brain's information processing, ignoring irrelevant information and reinforcing necessary information. AM uses a probability allocation method to assign higher attention (weight) to data points in the input time series data that have a significant impact on photovoltaic power prediction, highlighting the influence of important data information and thus improving the accuracy of the prediction model. The model structure of AM is as follows: Figure 4 As shown, xt (t∈[1,n]) is the input of BiLSTM, Qt (t∈[1,n]) is the hidden layer output obtained after each input passes through BiLSTM, and αt (t∈[1,n]) is the attention weight (probability assignment).

[0059] Weighting coefficient α t Calculate using equations (9)–(11)

[0060] e t = utanh(wQ t + b) (9)

[0061]

[0062] Where e t For Q t The attention probability allocation value is determined at time t, where u and w are weighting coefficients and b is the bias coefficient.

[0063] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A dynamic prediction system for the consumption of new energy sources in a regional power grid, characterized in that, include: The eigenvalue decomposition unit decomposes VMP data and photovoltaic power generation data into multiple scales through wavelet transform to obtain time series under each frequency component; The feature fusion unit is used to extract the spatial distribution features of the input data; BiLSTM units are used to mine the temporal correlation of input data sequences; The AM unit adaptively adjusts the weights of the BiLSTM model for each input data value according to the data order.

2. The regional power grid renewable energy consumption dynamic prediction system according to claim 1, characterized in that, The feature fusion unit employs a convolutional neural network.

3. The regional power grid renewable energy consumption dynamic prediction system according to claim 1, characterized in that, The BiLSTM unit includes an input layer, forward and reverse LSTM layers, forward and reverse LSTM layers, and an output layer.

4. The regional power grid renewable energy consumption dynamic prediction system according to claim 1, characterized in that, The weights of the attention units are: Among them, e t For Q t The attention probability allocation value is determined at time t, where u and w are weighting coefficients and b is the bias coefficient.

5. The regional power grid renewable energy consumption dynamic prediction system according to claim 1, characterized in that, Q t The attention probability allocation value e determined at time t t for: e t =utanh(wQ t +b)。

Images